ProjectTo{AbstractArray}(AbstractArray{NoTangent}) should be NoTangent

[oxinabox]

Right now we have

julia> ProjectTo([1.0, 2.0])([NoTangent(), NoTangent()])
2-element Vector{NoTangent}:
 NoTangent()
 NoTangent()

I propose we want:

julia> ProjectTo([1.0, 2.0])([NoTangent(), NoTangent()])
NoTangent()

They are semantically the same: they are bot hrepresetnations of a nondifferentiable zero.
But NoTangent is nicer.
I think this would be helpful for several cases,
and I think it would be nicely symmetric with ProjectTo(Array{Bool})=ProjectTo{NoTangent}()

Cases it would help with is things like higher order function such as map and sum etc
I think often you can avoid it outright by not even computing the element for the function object, but sometimes it comes up for hte args.
E.g. map(getindex, xs, inds) it would still benefit since it would compress those inds

It came up in ChainRulesTestUtils

[mcabbott]

Sounds good. Should ProjectTo([1.0, 2])(Any[NoTangent(), NoTangent()]) also collapse? Then it can't be done by dispatch, but can still be fast in the easy case:

julia> @btime all(dx -> dx isa NoTangent, $(Any[NoTangent() for _ in 1:10^6]))
  541.709 μs (0 allocations: 0 bytes)
true

julia> @btime all(dx -> dx isa NoTangent, $([NoTangent() for _ in 1:10^6]))
  0.875 ns (0 allocations: 0 bytes)
true

# Edit, also NB:

julia> @btime all(dx -> dx isa NoTangent, $(randn(10^6)))
  1.500 ns (0 allocations: 0 bytes)
false

[oxinabox]

I think that would make sense.
Things are always going to be slow if working with Array{Any}
this is a one time cost that will let us stop dealing with that cost, so it probably comes out in the wash

[mzgubic]

Closed by #424